On the identification of robot parameters by the classic calibration algorithms and error absorbing trees

In this paper, we propose a feasible method to construct a virtual manipulator in a 3-D graphics environment, which is equivalent to a real manipulator including its dynamics. For this purpose, we firstly calibrate parameters of robot dynamics by the classic two algorithms. One is to use a few special motions to identify viscous, dynamic and static frictions, masses of links, centers of the masses, mass matrixes of the links, and so on. The other is to use the least square method to identify robot parameters from a huge number of general motions. Unfortunately, both classic methods are not practically stable because each motion pattern includes noize and error. To overcome this drawback, we indirectly absorb position differences between experimental true and calculated virtual manipulators by three types of learning trees. Moreover, we directly memorize many neighbor dynamic motions by the same learning trees. As a result, when real and virtual robots are independently supervised by the PD control, their angular errors of rotational joints amount to zero. In addition, even though both robots are independently supervised as slave arms from a master arm by the same sequence of forces in a bilateral control based on the PD control, motion sequences of real and virtual slave arms equal to each other. As a result, a virtual manipulator can be used in a 3-D graphics animation, which is truly replaced of a real manipulator.